Statistical shape model creation method and device, equipment and medium
Through the combination of dynamic graph convolutional neural network and regularized flow model, end-to-end generation from CT point cloud to statistical shape model is achieved, solving the problem of traditional methods' lack of preprocessing dependence and nonlinear shape variation capture capabilities, and improving the accuracy of femoral head CT reconstruction and the applicability of clinical applications.
Patent Information
- Application Number
- CN202510450702.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-10
- Publication Date
- 2025-06-20
AI Technical Summary
The traditional statistical shape model construction method cannot be completed end-to-end, highly relies on the accuracy of pre-processing registration, and cannot effectively capture the nonlinear shape variability of the femoral head, limiting its effect in femoral head CT reconstruction and clinical applications.
Dynamic graph convolutional neural network (DGCNN) combined with regularized flow model is used to simplify the traditional multi-step preprocessing process through an end-to-end way from CT point cloud to statistical shape model. The sub-point cloud division is refined using anatomical key points to ensure that the generated model conforms to the anatomy of the femoral head.
It realizes high-precision end-to-end generation from CT point cloud to statistical shape models, improves the accuracy of femoral head CT reconstruction and the applicability of clinical applications, and provides reliable tools to support orthopedic surgical planning and diagnosis.
Smart Images
Figure CN120182245A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical-engineering integration, and particularly to a method, device, equipment and medium for creating a statistical shape model of the femoral head based on flow. Background Art
[0002] Statistical Shape Model (SSM) is widely used in medical image processing for organ segmentation and morphological analysis, and is particularly significant in the CT analysis of the femoral head.
[0003] The construction of traditional statistical shape models usually uses the Principal Component Analysis (PCA) method. PCA is a dimensionality reduction technique that projects high-dimensional data into a low-dimensional space, retaining the most important variability information in the data. In statistical shape models, PCA is used to identify the main variation directions of the shape (referred to as principal components or eigenvectors), and these principal components are used to approximate the variability of the shape.
[0004] However, the method for constructing a statistical shape model based on PCA requires complex preprocessing steps (such as registration and establishment of corresponding points) and cannot achieve end-to-end creation. Moreover, this method highly depends on the accuracy of preprocessing registration. If the registration error is large, the model quality will decrease significantly. Most importantly, the shape changes of the femur contain non-linear features, such as bending, twisting or local deformation, etc. This non-linear variability is particularly obvious in pathological conditions (such as femoral head necrosis or dysplasia). Since PCA cannot effectively capture these non-linear features, the construction of traditional statistical shape models for the femoral head statistical shape model is restricted. The principal components extracted by PCA are linear combinations of the original features, losing the direct meaning of the original features. In the femoral head shape modeling, these principal components may not correspond to specific anatomical features, such as the sphericity or neck-shaft angle of the femoral head. This makes the interpretability of the model poor and difficult to be directly applied to clinical scenarios (such as diagnosis or surgical planning).
[0005] It can be seen that the shape of the femoral head is complex and variable, and it is difficult for traditional methods to accurately capture its anatomical features, which limits clinical applications. Summary of the Invention
[0006] In view of the above problems, the present invention provides a method, device, equipment and medium for creating a statistical shape model to overcome or at least partially solve the above problems. The aim is to realize the construction of the femoral head shape in an end-to-end manner. It solves the problems that the existing SSM creation method cannot be completed end-to-end and has a strong dependence on the accuracy of preprocessing registration, resulting in poor CT reconstruction effect of the femoral head, etc.
[0007] The present invention provides the following solutions:
[0008] A method for creating a statistical shape model, comprising:
[0009] Using image segmentation technology to separate a three-dimensional model of the femoral head from the original femoral head CT image and extracting an initial three-dimensional point cloud of the femoral head from the three-dimensional model of the femoral head;
[0010] Using a key point extraction module to extract anatomical key points from the original femoral head CT image;
[0011] Combining the initial three-dimensional point cloud, using a dynamic graph convolutional neural network to extract multi-scale local geometric features of the point cloud, constructing a dynamic adjacency graph of the point cloud through the K-nearest neighbor algorithm, and using the EdgeConv operation to extract local and global geometric features to obtain a point cloud feature vector;
[0012] Mapping the point cloud feature vector to a sub-structure category probability space through a fully connected layer; embedding the anatomical key points into the feature space as a prior constraint for sub-point cloud division, and combining the Gumbel-Softmax technique for differentiable sampling to obtain a point cloud label for each point, so as to divide the initial three-dimensional point cloud into multiple sub-point clouds, and the point cloud label includes the sub-structure category label to which each point belongs;
[0013] Applying a regularization flow model to each of the sub-point clouds, so that the regularization flow model generates a target point cloud distribution from a noise distribution through a reversible transformation, and generates a femoral head statistical shape model that conforms to the anatomical structure with the point cloud label as a condition.
[0014] Preferably: after feature extraction, a variational auto-encoding mechanism is introduced to generate a mean vector and a variance vector of the global feature of the point cloud, and a resampling technique is used to sample latent variables from a Gaussian distribution.
[0015] Preferably: using the EdgeConv operation to extract local and global geometric features to obtain a point cloud feature vector includes:
[0016] Performing a convolution operation on the dynamic adjacency graph to learn the edge features between points and neighbors;
[0017] Calculating edge features for each point and its neighbor points;
[0018] Aggregating the edge features of each point to generate a new feature representation;
[0019] Extracting higher-level features by stacking multiple layers of DGCNN;
[0020] Repeating the construction of the dynamic adjacency graph and the EdgeConv operation, and updating the adjacency graph layer by layer based on the feature space of the previous layer;
[0021] Aggregate local features into global features to capture the overall information of the point cloud;
[0022] Output the point cloud feature vector according to the task requirements.
[0023] Preferably: Embedding the anatomical key points into the feature space, and the prior constraints for sub-point cloud division include:
[0024] Calculate the weighting factor through the Euclidean distance or direction vector between the anatomical key points and the point cloud feature vector, and adjust the probability distribution of each point belonging to the sub-structure.
[0025] Preferably: The anatomical key points at least include the center of the femoral head and the intersection of the femoral neck axis.
[0026] Preferably: Pre-define the sub-structure category to which each anatomical key point belongs during the training process of sub-point cloud division; add key point classification loss to the loss function to ensure that the key points and their adjacent areas are correctly assigned to the corresponding sub-structures.
[0027] Preferably: Optimize the log-likelihood loss during the training of the normalizing flow model to ensure the geometric accuracy and distribution consistency of the generated point cloud.
[0028] A statistical shape model creation device for performing the above-mentioned statistical shape model creation method, the device includes:
[0029] An initial point cloud acquisition unit for separating a three-dimensional model of the femoral head from the original femoral head CT image by using image segmentation technology and extracting an initial three-dimensional point cloud of the femoral head from the three-dimensional model of the femoral head;
[0030] An anatomical key point acquisition unit for extracting anatomical key points from the original femoral head CT image by using a key point extraction module;
[0031] A point cloud feature acquisition unit for combining the initial three-dimensional point cloud to extract multi-scale local geometric features of the point cloud by using a dynamic graph convolutional neural network, constructing a dynamic adjacency graph of the point cloud through the K-nearest neighbor algorithm, and extracting local and global geometric features by using EdgeConv operations to obtain a point cloud feature vector;
[0032] A sub-point cloud division unit for mapping the point cloud feature vector to the sub-structure category probability space through a fully connected layer; embedding the anatomical key points into the feature space, using the Gumbel-Softmax technique as the prior constraint for sub-point cloud division to perform differentiable sampling to obtain the point cloud labels of each point, so as to divide the initial three-dimensional point cloud into multiple sub-point clouds, and the point cloud labels include the sub-structure category labels to which each point belongs;
[0033] A model generation unit for applying a normalizing flow model to each of the sub-point clouds, so that the normalizing flow model generates a target point cloud distribution from a noise distribution through a reversible transformation and generates a femoral head statistical shape model conforming to the anatomical structure conditioned on the point cloud label.
[0034] A statistical shape model creation device, the device includes a processor and a memory:
[0035] The memory is used to store program code and transmit the program code to the processor;
[0036] The processor is used to execute the above-mentioned statistical shape model creation method according to the instructions in the program code.
[0037] A computer-readable storage medium for storing program code for executing the above-mentioned statistical shape model creation method.
[0038] According to the specific embodiments provided by the present invention, the following technical effects are disclosed:
[0039] A statistical shape model creation method, device, device and medium provided by an embodiment of the present application, through the combination of a DGCNN network and a normalizing flow, realizes the end-to-end generation from CT point cloud to a statistical shape model, simplifying the traditional multi-step preprocessing process. Using anatomical key points to refine the sub-point cloud division ensures that the division result conforms to the femoral head anatomical structure and improves clinical applicability. The normalizing flow model generates high-precision sub-point clouds through reversible transformation, adapting to the complexity and variability of the femoral head morphology. Experiments show that this method has higher accuracy than traditional SSM methods in femoral head CT reconstruction, providing a reliable tool for orthopedic surgical planning and diagnosis.
[0040] Of course, it is not necessary for any product implementing the present invention to achieve all the above advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.
[0042] Figure 1 is a flowchart of the statistical shape model creation method provided by an embodiment of the present invention;
[0043] Figure 2 is an overall flowchart block diagram provided by an embodiment of the present invention;
[0044] Figure 3 is the principle flowchart of the dynamic graph convolution module provided by the embodiment of the present invention;
[0045] Figure 4 is the workflow diagram of the sub-point cloud division module provided by the embodiment of the present invention;
[0046] Figure 5 is the schematic diagram of the statistical shape model creation device provided by the embodiment of the present invention;
[0047] Figure 6 is the schematic diagram of the statistical shape model creation device provided by the embodiment of the present invention. Detailed implementation manners
[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention belong to the scope of protection of the present invention.
[0049] Refer to Figure 1 , a statistical shape model creation method provided by the embodiment of the present invention. As Figure 1 shown, the method may include:
[0050] S101: Use image segmentation technology to separate the three-dimensional model of the femoral head from the original femoral head CT image and extract the initial three-dimensional point cloud of the femoral head from the three-dimensional model of the femoral head;
[0051] S102: Use the key point extraction module to extract anatomical key points from the original femoral head CT image; Specifically, in the embodiment of the present application, the anatomical key points may include at least the center of the femoral head, the intersection of the femoral neck axis, etc.
[0052] S103: Combine the initial three-dimensional point cloud and use a dynamic graph convolutional neural network to extract the multi-scale local geometric features of the point cloud, construct a dynamic adjacency graph of the point cloud through the K-nearest neighbor algorithm, and use the EdgeConv operation to extract local and global geometric features to obtain the point cloud feature vector; Further, in the embodiment of the present application, after feature extraction, a variational auto-encoding mechanism may be introduced to generate the mean vector and variance vector of the point cloud global features, and a resampling technology may be used to sample latent variables from the Gaussian distribution.
[0053] Using the EdgeConv operation to extract local and global geometric features to obtain the point cloud feature vector includes:
[0054] Perform a convolution operation on the dynamic adjacency graph to learn the edge features between points and neighbors;
[0055] Calculate edge features for each point and its neighboring points;
[0056] Aggregate the edge features of each point to generate a new feature representation;
[0057] Extract higher-level features by stacking multiple layers of DGCNN;
[0058] Repeat the construction of the dynamic adjacency graph and EdgeConv operations, and update the adjacency graph for each layer based on the feature space of the previous layer;
[0059] Aggregate local features into global features to capture the overall information of the point cloud;
[0060] Output the point cloud feature vector according to the task requirements.
[0061] S104: Map the point cloud feature vector to the sub-structure category probability space through a fully connected layer; Embed the anatomical key points into the feature space as a prior constraint for sub-point cloud division, and combine the Gumbel-Softmax technique to perform differentiable sampling to obtain the point cloud labels of each point, so as to divide the initial three-dimensional point cloud into multiple sub-point clouds. The point cloud labels include the sub-structure category labels to which each point belongs; Specifically, in implementation, the embodiments of the present application may provide embedding the anatomical key points into the feature space as a prior constraint for sub-point cloud division, including:
[0062] Calculate the weighting factor through the Euclidean distance or direction vector between the anatomical key points and the point cloud feature vector, and adjust the probability distribution of each point belonging to the sub-structure.
[0063] Further, pre-define the sub-structure category to which each anatomical key point belongs during the training process of sub-point cloud division; Add key point classification loss to the loss function to ensure that the key points and their neighboring regions are correctly assigned to the corresponding sub-structures.
[0064] S105: Apply a normalizing flow model to each sub-point cloud, so that the normalizing flow model generates the target point cloud distribution from the noise distribution through a reversible transformation, and generates a statistical shape model of the femoral head that conforms to the anatomical structure with the point cloud label as a condition. Specifically, in implementation, the embodiments of the present application may provide optimizing the log-likelihood loss during the training of the normalizing flow model to ensure the geometric accuracy and distribution consistency of the generated point cloud.
[0065] The statistical shape model creation method provided by the embodiments of the present application uses dynamic graph convolution to directly extract features from the point cloud dataset, and uses an attention mechanism to fuse the extracted features to achieve learning of population anatomical features. The overall point cloud is divided into multiple sub-point clouds for separate reconstruction. The number of sub-point clouds is adaptively matched through a label mask. Through normalizing flow, a three-dimensional shape that conforms to the population anatomical feature distribution is reconstructed. In addition, this method can learn from unregistered datasets and does not require the construction of point-by-point correspondences, greatly simplifying the process of constructing a statistical shape model.
[0066] The method provided by the embodiments of the present application will be introduced in detail below.
[0067] The method provided by the embodiments of the present application uses a dynamic graph convolutional neural network (DGCNN) to replace the traditional PointNet for extracting multi-scale local geometric features of the point cloud. A dynamic adjacency graph between points is constructed through the k-nearest neighbor algorithm (KNN), and combined with the EdgeConv convolution operation to learn the structural relationship between points.
[0068] After feature extraction, a variational auto-encoding mechanism is introduced to generate the mean vector and variance vector of the global features of the point cloud. A resampling technique is used to sample the latent variable s_X from the Gaussian distribution to improve the model's ability to model the variability of the point cloud and its generalization ability.
[0069] Taking the DGCNN sampled features as input, it is mapped to the class probability space through a fully connected layer. The Gumbel-Softmax technique is introduced to achieve differentiable sampling of sub-structure categories, enabling the entire partitioning process to be trained end-to-end. The sub-structure category label (one-hot) to which each point belongs is output, supporting both hard partitioning and soft partitioning. The overall point cloud is divided into multiple sub-point clouds.
[0070] During the partitioning process, anatomical key points of the femoral head (such as the center of the femoral head, the intersection of the femoral neck axis, etc.) are predefined, and the three-dimensional coordinates of these key points are obtained through medical image analysis or expert annotation. The key points are embedded into the feature space as a prior constraint for sub-point cloud partitioning.
[0071] Specifically, before Gumbel-Softmax sampling, a weighting factor is calculated through the Euclidean distance or direction vector between the anatomical key points and the point cloud feature vectors to adjust the probability distribution of each point belonging to a sub-structure. Through key point constraints, it is ensured that the boundaries of the sub-point clouds are consistent with the anatomical regions of the femoral head (such as the head, neck), enhancing the anatomical significance and consistency of the partitioning.
[0072] The features of the sub-point cloud are input into the normalizing flow model, and the noise distribution is mapped to the target point cloud distribution through a series of invertible transformations. In the specific implementation, the Glow normalizing flow architecture is adopted, and the log-likelihood loss is optimized during training to ensure the geometric accuracy and distribution consistency of the generated point cloud. To adapt to the complexity of the femoral head point cloud, conditional inputs (sub-point cloud class labels) are introduced to guide the flow model to generate point clouds that conform to the anatomical structure. The generated sub-point clouds are aligned with adjacent sub-point clouds through coordinate transformation to form a seamlessly connected local structure.
[0073] As Figure 2 shown, the overall process includes:
[0074] 1. Data preprocessing: The femoral head region is separated through image segmentation technology, and denoising processing is applied to reduce noise interference and extract the three-dimensional point cloud data of the femoral head.
[0075] 2. Anatomical key points are extracted from the original femoral head CT image using the key point extraction module
[0076] 3. Feature extraction: Using the Dynamic Graph Convolutional Neural Network (DGCNN), a dynamic adjacency graph of the point cloud is constructed through the K-Nearest Neighbor (KNN) algorithm, and local and global geometric features are extracted using the EdgeConv operation to capture the topological structure and anatomical characteristics of the femoral head.
[0077] 4. Sub-point cloud division: The feature vectors are mapped to the sub-structure category probability space through a fully connected layer. Combining the Gumbel-Softmax technique for differentiable sampling and using the anatomical key points as constraints, the division probability of the sub-point cloud is optimized.
[0078] 5. Sub-point cloud generation: Apply the normalizing flow model (Glow) to each sub-point cloud, generate the target point cloud distribution from the noise distribution through invertible transformations, and embed the class label as a condition to ensure that the generated result conforms to the anatomical structure.
[0079] Combined Figure 3 , the working principle of the Dynamic Graph Convolutional Neural Network (DGCNN) module provided in this application is described in detail.
[0080] 1. Input points.
[0081] The process starts with the initial point cloud data generated after preprocessing the femoral head CT image.
[0082] The input point cloud P contains N points, and each point is represented by three-dimensional coordinates (x, y, z).
[0083] 2. Construct a dynamic adjacency graph.
[0084] Construct a local adjacency relationship for each point to capture the topological structure of the point cloud. Use the K-Nearest Neighbors (KNN) algorithm for each point to find its K nearest neighbor points and generate a dynamic adjacency graph.
[0085] 3. EdgeConv operation.
[0086] Perform a convolution operation on the adjacency graph to learn the edge features between points and their neighbors.
[0087] Calculate the edge features for each point and its neighbor points.
[0088] Aggregate (e.g., max pooling) the edge features for each point to generate a new feature representation.
[0089] 4. Multi-layer feature extraction.
[0090] Extract higher-level features by stacking multiple layers of DGCNN.
[0091] Repeat the "construction of the dynamic adjacency graph" and the "EdgeConv operation", and update the adjacency graph for each layer based on the feature space of the previous layer.
[0092] 5. Global feature aggregation.
[0093] Aggregate the local features into global features to capture the overall information of the point cloud.
[0094] 6. Output features.
[0095] Output the final point cloud feature vector according to the task requirements.
[0096] Combined with Figure 4 , the implementation process of refining the sub-point cloud division by using anatomical key points is as follows:
[0097] 1. Key point embedding.
[0098] Embed the spatial information of the anatomical key points into the point cloud features as the prior guidance for the division.
[0099] 2. Feature fusion. Fuse the original point cloud features with the key point influence vectors to generate enhanced features.
[0100] 3. Sub-structure probability mapping. Map the enhanced features to the probability space of sub-structure categories.
[0101] 4. Gumbel-Softmax sampling. Implement differentiable class assignment through the Gumbel-Softmax technique.
[0102] 5. Key point constraint optimization. During the training process, use the known class information of the key points to optimize the division results.
[0103] Pre-define the sub-structure category to which each key point belongs (for example, the center of the femoral head belongs to the head sub-point cloud). Add the key point classification loss to the loss function. Ensure that the key points and their adjacent regions are correctly assigned to the corresponding sub-structures.
[0104] It can be seen that the embodiment of the present application provides an innovative method for reconstructing the CT sub-point cloud of the femoral head. By integrating the Dynamic Graph Convolutional Neural Network (DGCNN), anatomical key point constraints, and the regularization flow model, an efficient and accurate end-to-end femoral head statistical shape model creation is achieved. Its core advantages are as follows: using DGCNN to capture the local topology and geometric features of the point cloud, refining the sub-point cloud division through anatomical key points, and combining the regularization flow to generate a high-precision point cloud adapted to individual variations. Finally, the entire reconstruction process is seamlessly integrated, significantly improving the accuracy and robustness of 3D reconstruction, and providing important support for orthopedic surgical planning and diagnosis.
[0105] In summary, the statistical shape model creation method provided by the present application realizes the end-to-end generation from the CT point cloud to the statistical shape model through the combination of the DGCNN network and the Gumbel-Softmax technology, simplifying the traditional multi-step preprocessing process. Using anatomical key points to refine the sub-point cloud division ensures that the division results conform to the anatomical structure of the femoral head and improves clinical applicability. The regularization flow model generates a high-precision sub-point cloud through reversible transformation, adapting to the complexity and variability of the femoral head morphology. Experiments show that this method has higher accuracy than traditional SSM methods in femoral head CT reconstruction, providing a reliable tool for orthopedic surgical planning and diagnosis.
[0106] See Figure 5 , the embodiment of the present application can also provide a statistical shape model creation device, such as Figure 5 shown, for executing the above-mentioned statistical shape model creation method. The device may include:
[0107] An initial point cloud acquisition unit 501, configured to separate a three-dimensional model of the femoral head from the original femoral head CT image by using image segmentation technology and extract an initial three-dimensional point cloud of the femoral head from the three-dimensional model of the femoral head;
[0108] An anatomical key point acquisition unit 502, configured to extract anatomical key points from the original femoral head CT image by using a key point extraction module;
[0109] A point cloud feature acquisition unit 503, configured to extract multi-scale local geometric features of the point cloud by using a dynamic graph convolutional neural network in combination with the initial three-dimensional point cloud, construct a dynamic adjacency graph of the point cloud through the K-nearest neighbor algorithm, and extract local and global geometric features by using EdgeConv operations to obtain a point cloud feature vector;
[0110] The sub - point cloud division unit 504 is used to map the point cloud feature vector to the sub - structure category probability space through a fully - connected layer; embed the anatomical key points into the feature space as the prior constraint for sub - point cloud division, and perform differentiable sampling by combining the Gumbel - Softmax technique to obtain the point cloud labels of each point, so as to divide the initial three - dimensional point cloud into multiple sub - point clouds, where the point cloud labels include the sub - structure category labels to which each point belongs.
[0111] The model generation unit 505 is used to apply a normalizing flow model to each of the sub - point clouds, so that the normalizing flow model generates the target point cloud distribution from the noise distribution through an invertible transformation, and generates a femoral head statistical shape model that conforms to the anatomical structure with the point cloud labels as the condition.
[0112] An embodiment of the present application can also provide a statistical shape model creation device, and the device includes a processor and a memory:
[0113] The memory is used to store program code and transmit the program code to the processor;
[0114] The processor is used to execute the steps of the above - mentioned statistical shape model creation method according to the instructions in the program code.
[0115] As Figure 6 shown, a statistical shape model creation device provided by an embodiment of the present application may include: a processor 10, a memory 11, a communication interface 12, and a communication bus 13. The processor 10, the memory 11, and the communication interface 12 all complete communication with each other through the communication bus 13.
[0116] In an embodiment of the present application, the processor 10 may be a central processing unit (CPU), an application - specific integrated circuit, a digital signal processor, a field - programmable gate array, or other programmable logic devices, etc.
[0117] The processor 10 may call the program stored in the memory 11. Specifically, the processor 10 may execute the operations in the embodiment of the statistical shape model creation method.
[0118] The memory 11 is used to store one or more programs, and the program may include program code. The program code includes computer operation instructions. In an embodiment of the present application, the memory 11 stores at least a program for implementing the following functions:
[0119] Using image segmentation technology, a three - dimensional model of the femoral head is separated from the original femoral head CT image, and an initial three - dimensional point cloud of the femoral head is extracted from the three - dimensional model of the femoral head;
[0120] The key point extraction module is used to extract anatomical key points from the original femoral head CT images;
[0121] Combined with the initial three-dimensional point cloud, a dynamic graph convolutional neural network is used to extract multi-scale local geometric features of the point cloud. A dynamic adjacency graph of the point cloud is constructed through the K-nearest neighbor algorithm, and local and global geometric features are extracted using EdgeConv operations to obtain a point cloud feature vector;
[0122] The point cloud feature vector is mapped to the sub-structure category probability space through a fully connected layer; the anatomical key points are embedded into the feature space as prior constraints for sub-point cloud division, and combined with the Gumbel-Softmax technique for differentiable sampling to obtain the point cloud labels of each point, so as to divide the initial three-dimensional point cloud into multiple sub-point clouds. The point cloud labels include the sub-structure category labels to which each point belongs;
[0123] Apply a normalizing flow model to each of the sub-point clouds, so that the normalizing flow model generates a target point cloud distribution from a noise distribution through an invertible transformation, and generates a femoral head statistical shape model that conforms to the anatomical structure with the point cloud labels as conditions. In a possible implementation, the memory 11 may include a program storage area and a data storage area. Among them, the program storage area may store an operating system and application programs required for at least one function (such as a file creation function and a data reading and writing function), etc.; the data storage area may store data created during use, such as initialization data, etc.
[0124] In addition, the memory 11 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device or other volatile solid-state storage devices.
[0125] The communication interface 12 may be an interface of a communication module for connecting to other devices or systems.
[0126] Of course, it should be noted that Figure 6 The structure shown does not constitute a limitation on the statistical shape model creation device in the embodiments of the present application. In practical applications, the statistical shape model creation device may include more or fewer components than Figure 6 shown, or combine some components.
[0127] The embodiments of the present application may also provide a computer-readable storage medium for storing program codes for executing the steps of the above-mentioned statistical shape model creation method.
[0128] It should be noted that in this text, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element.
[0129] From the description of the above embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution of this application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.
[0130] Each embodiment in this specification is described in a progressive manner. For the same or similar parts between the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for a system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the method embodiment. The systems and system embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they can be located in one place, or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.
[0131] The above is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.
Claims
1. A method for creating a statistical shape model, characterized in that: include: Using image segmentation technology to separate the original femoral head CT image to obtain a femoral head three-dimensional model and extracting an initial three-dimensional point cloud of the femoral head from the femoral head three-dimensional model; The key point extraction module is used to extract anatomical key points from the original femoral head CT image; A dynamic graph convolutional neural network is used to extract multi-scale local geometric features of the point cloud in combination with the initial three-dimensional point cloud, a dynamic adjacency graph of the point cloud is constructed by a K-nearest neighbor algorithm, and an EdgeConv operation is used to extract local and global geometric features to obtain a point cloud feature vector; The point cloud feature vector is mapped to the substructure category probability space through a fully connected layer; the anatomical key points are embedded in the feature space as a priori constraints for the sub-point cloud division, and the point cloud labels of each point are obtained by differentiable sampling in combination with the Gumbel-Softmax technology, so as to divide the initial three-dimensional point cloud into multiple sub-point clouds, and the point cloud labels include the substructure category labels to which each point belongs; A regularized flow model is applied to each of the sub-point clouds so that the regularized flow model generates a target point cloud distribution from a noise distribution through a reversible transformation, and a femoral head statistical shape model that conforms to the anatomical structure is generated using the point cloud label as a condition.
2. The statistical shape model creation method according to claim 1, characterized in that: After feature extraction, the variational autoencoder mechanism is introduced to generate the mean vector and variance vector of the global features of the point cloud, and the resampling technique is used to sample the latent variables from the Gaussian distribution.
3. The statistical shape model creation method according to claim 1, characterized in that: Using EdgeConv operation to extract local and global geometric features to obtain point cloud feature vectors includes: Performing a convolution operation on the dynamic adjacency graph to learn edge features between points and neighbors; Calculate edge features for each point and its neighboring points; Aggregate the edge features of each point to generate a new feature representation; By stacking multiple layers of DGCNN, more advanced features can be extracted; Repeatedly construct the dynamic adjacency graph and EdgeConv operations, and each layer updates the adjacency graph based on the feature space of the previous layer; Aggregate local features into global features to capture the overall information of the point cloud; The point cloud feature vector is output according to task requirements.
4. The statistical shape model creation method according to claim 1, characterized in that: The anatomical key points are embedded in the feature space as the prior constraints for the sub-point cloud segmentation include: The weighting factor is calculated by the Euclidean distance or direction vector between the anatomical key point and the point cloud feature vector, and the probability distribution of the substructure to which each point belongs is adjusted.
5. The statistical shape model creation method according to claim 4, characterized in that: The anatomical key points include at least the center of the femoral head and the intersection of the femoral neck axis.
6. The statistical shape model creation method according to claim 5, characterized in that: During the sub-point cloud segmentation training process, the substructure category to which each anatomical key point belongs is predefined; and the key point classification loss is added to the loss function to ensure that the key point and its neighboring area are correctly assigned to the corresponding substructure.
7. The statistical shape model creation method according to claim 1, characterized in that: The regularized flow model optimizes the log-likelihood loss during training to ensure the geometric accuracy and distribution consistency of the generated point cloud.
8. A statistical shape model creation device, characterized in that: Used to execute the statistical shape model creation method according to any one of claims 1 to 7, the device comprising: An initial point cloud acquisition unit is used to separate and obtain a femoral head three-dimensional model from the original femoral head CT image using an image segmentation technology and extract an initial three-dimensional point cloud of the femoral head from the femoral head three-dimensional model; An anatomical key point acquisition unit is used to extract anatomical key points from the original femoral head CT image using a key point extraction module; A point cloud feature acquisition unit is used to extract multi-scale local geometric features of the point cloud using a dynamic graph convolutional neural network in combination with the initial three-dimensional point cloud, construct a dynamic adjacency graph of the point cloud using a K-nearest neighbor algorithm, and extract local and global geometric features using an EdgeConv operation to obtain a point cloud feature vector; A sub-point cloud division unit is used to map the point cloud feature vector to a sub-structure category probability space through a fully connected layer; embed the anatomical key points into the feature space as a priori constraints for sub-point cloud division, and perform differentiable sampling in combination with the Gumbel-Softmax technology to obtain point cloud labels for each point, so as to divide the initial three-dimensional point cloud into multiple sub-point clouds, wherein the point cloud labels include the sub-structure category labels to which each point belongs; The model generation unit is used to apply a regularized flow model to each of the sub-point clouds so that the regularized flow model generates a target point cloud distribution from a noise distribution through a reversible transformation, and generates a femoral head statistical shape model that conforms to the anatomical structure using the point cloud label as a condition.
9. A statistical shape model creation device, characterized in that: The device comprises a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the statistical shape model creation method according to any one of claims 1-7 according to the instructions in the program code.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store program codes, and the program codes are used to execute the statistical shape model creation method according to any one of claims 1 to 7.